Decide where to put sampler recommender

[chaozg]

It is concluded from our stand-up that:

• Indeed, the above suggestion from @jakobsj is neat and intuitive;
• @amal-ghamdi mentioned that, however, implementing such a feature requires direct changes to density/distribution, which we have deliberately avoided so far. Currently, we do sampling via the interface of BayesianProblem or sampler, instead of with density/distribution directly. Introducing this feature will break this consistency and convention.
• recommender does not seem to fit well at the very high level of cuqi, cuqi.utilities or cuqi.sampler.
• We agree that more careful consideration is needed on this issue.
• Therefore, it's decided that we put recommender back to the experimental module, which is appropriate in the sense that the feature is still experimental.

Originally posted by @chaozg in #693 (comment)